@ashkitten: Thanks for taking a look at this, I updated the package to 2020.Q1.3 and the hash should be correct now.

when i set environment.variables.VK_ICD_FILENAMES = "${pkgs.amdvlk}/share/vulkan/icd.d/amd_icd64.json" and reboot, running vulkaninfo gives me a segfault, as well as vkcube. do these only work with mesa, or is there something wrong with my setup?

vkcube and vulkaninfo both work with amdvlk and there is nothing particularly wrong with your setup but there are a few weird errors.

If $DISPLAY is not set, vulkaninfo works fine with amdvlk and radv.
If $DISPLAY is set, vulkaninfo only works with amdvlk if the following packages are in $LD_LIBRARY_PATH: xorg.libX11.
If $DISPLAY is set, vkcube only works with amdvlk if the following packages are in $LD_LIBRARY_PATH: xorg.libX11, xorg.libxcb, xorg.libxshmfence.
$WAYLAND_DISPLAY is ignored by both, they only work with xwayland.

I also tested the code of a simple Vulkan tutorial with SDL (with Wayland support).
If $WAYLAND_DISPLAY is set, works out-of-the-box with amdvlk and radv.
If only $DISPLAY is set on xwayland, it hangs somewhere when presenting the swapchain, happens for amdvlk and radv.

I added these libraries to the rpath, so vkcube and vulkaninfo should now work without any extra configuration.

interesting that they work fine with radv anyways

can confirm that it works now, by the way

just tried to build this on my laptop and it says that the sha256 is invalid again. i'm not sure how repo projects work, but the hash seems to be unstable which is an issue.

Thank you for all the testing!
You are right and I now see why. The repository that contains the repo manifest also is included in the repo: https://github.com/GPUOpen-Drivers/AMDVLK/blob/v-2020.Q1.3/default.xml#L16
But as it is hard to reference itself, it just references master, which of course changes everytime a new release gets pushed.

Possibilities to fix this are

1. Extend fetchRepoProject to support overriding fetched repositories with a fixed revision, this would allow us to pin the AMDVLK repository so the hash should not change anymore.
2. Convince upstream to do two commits for every release and reference the first commit in the default.xml so we can get reproducible builds.
3. Push our own default.xml file to nixpkgs which is based on the upstream one but contains a fixed hash for AMDVLK.

I think 3 is the easiest way to fix this. @ashkitten, what do you think?

that sounds reasonable.

Many releases later, we are now at 2020.Q2.3 and the package declaration got slimmer 🎉.
Can someone maybe merge this?

Updated to 2020.Q2.4.
@Ma27, could you or someone else maybe merge this please? Thanks!

Don't have the power to merge anything I'm afraid, but I just wanted to note that I tried this just now to play with the new Path of Exile Vulkan beta (RADV crashes) and it worked wonderfully.

Can also confirm this works for me

Perhaps there should be a module that sets it up correctly?

Thanks for trying it out!

Perhaps there should be a module that sets it up correctly?

Putting the packages into environment.systemPackages makes it the default vulkan driver for me, is that what you mean?

I didn't actually try that and just went straight to setting the environment variable, I was thinking of a module that would set that environment variable for you but I guess if adding it to systemPackages works then this is not necessary

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/packaging-amdvlk/4795/5

Hello,

Great work, this works well on my R9 390.
vulkan-loader will find amd_icd64.json if I simply add the package to environment.systemPackages, but it won't pick it, supposedly because it finds stuff in /run/opengl-driver first.

I tried a few things and while it doesn’t change anything, adding amdvlk to hardware.opengl.extraPackages seems like a cleaner way than environment.systemPackages.
Adding it to any of those means both, radv/mesa and amdvlk will be found and exposed as possible GPUs. vulkaninfo | grep GPU (part of vulkan-tools) looks like this for me:

…
GPU id : 0 (AMD Radeon (TM) RX 480 Graphics):
GPU id : 1 (AMD RADV POLARIS10 (LLVM 9.0.1)):
GPU id : 0 (AMD Radeon (TM) RX 480 Graphics):
GPU id : 1 (AMD RADV POLARIS10 (LLVM 9.0.1)):
…

Every application decides itself which GPU/which driver it uses or it allows the user to choose. To force a certain driver, e.g. amdvlk or radv, setting

# amdvlk
VK_ICD_FILENAMES=/run/opengl-driver/share/vulkan/icd.d/amd_icd64.json
# or radv
VK_ICD_FILENAMES=/run/opengl-driver/share/vulkan/icd.d/radeon_icd.x86_64.json

can be used. If amdvlk is in environment.systemPackages instead of hardware.opengl.extraPackages it would be VK_ICD_FILENAMES=/run/current-system/sw/share/vulkan/icd.d/amd_icd64.json.
Running env VK_ICD_FILENAMES=/run/opengl-driver/share/vulkan/icd.d/amd_icd64.json vulkaninfo | grep GPU can be used to confirm the driver(s) in use.

Working well on my system too.
I'm using environment.variables.VK_ICD_FILENAMES = "${amdvlk}/share/vulkan/icd.d/amd_icd64.json";

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/prs-already-reviewed/2617/173

Also, tested with

nixpkgs-review rev HEAD
$ export VK_ICD_FILENAMES=results/amdvlk/share/vulkan/icd.d/amd_icd64.json
$ export DISPLAY=:1
$ vkcube

and it still works.

so, to enable the vulkan drivers, you just need to add amdvlk to hardware.opengl.extraPackages and then it's done ?
does this work with both radeon and amdgpu drivers ?

so, to enable the vulkan drivers, you just need to add amdvlk to hardware.opengl.extraPackages and then it's done ?

You don't actually have to do anything to use Vulkan drivers ;). If you have hardware.opengl.enable = true;, the Mesa Vulkan driver for amdgpu will be used. By putting amdvlk in extraPackages, you add this driver as an additional option.

I think both, RADV (the default AMD Vulkan driver which is part of mesa and installed when you set hardware.opengl.enable = true;) and amdvlk need the amdgpu kernel module, i.e. they do not work on pre-GCN GPUs where only the radeon kernel module is supported.
For early GCN GPUs, the amdgpu driver has to be explicitly enabled: https://wiki.archlinux.org/index.php/AMDGPU#Enable_Southern_Islands_(SI)_and_Sea_Islands_(CIK)_support